Authors:
Leonidas Alagialoglou
1
;
Ioannis Manakos
2
;
Olga Brovkina
3
;
Jan Novotný
3
and
Anastasios Delopoulos
1
Affiliations:
1
Multimedia Understanding Group, Aristotle University of Thessaloniki, Greece
;
2
Information Technologies Institute, Centre for Research and Technology Hellas, Thessaloniki, Greece
;
3
Department of Remote Sensing, Global Change Research Institute of the Czech Academy of Sciences, CzechGlobe, Brno, Czech Republic
Keyword(s):
Canopy Height Estimation, Deep Learning, Fine-Tuning, Forest, Sentinel-2, Data-Centric AI, Uncertainty Estimation, Tree Species, Airborne Laser Scanning.
Abstract:
This study evaluates the performance of a lightweight convolutional Long Short-Term Memory (ConvLSTM)based deep learning model for estimating canopy height across three test areas in the Czech Republic using Sentinel-2 time series data. The model, initially trained on forest data from Germany and Switzerland, incorporate uncertainty quantification techniques and was fine-tuned and evaluated using dense airborne laser scanning (ALS) data collected between 2022 and 2024. Results show that fine-tuning reduced mean absolute error (MAE) from 4.26 m to 2.74 m in the primary test area, with similar improvements across other regions. Species-specific uncertainties were also analyzed, highlighting performance variations between deciduous and coniferous forests.